We describe the 'Structure-Activity Relationship (SAR) Matrix' (SARM) methodology that is based upon a special two-step application of the matched molecular pair (MMP) formalism. The SARM method has originally been designed for the extraction, organization, and visualization of compound series and associated SAR information from compound data sets. It has been further developed and adapted for other applications including compound design, activity prediction, library extension, and the navigation of multi-target activity spaces. The SARM approach and its extensions are presented here in context to introduce different types of applications and provide an example for the evolution of a computational methodology in pharmaceutical research.

Analog series were systematically extracted from more than 650 000 bioactive compounds originating from medicinal chemistry and screening sources and more than 3.6 million commercial compounds that were not biologically annotated. Then, analog series-based (ASB) scaffolds were generated. For each scaffold from a bioactive series, a target profile was derived and ASB scaffolds shared by bioactive and commercial compounds were determined. On the basis of our analysis, large segments of commercial chemical space were not yet explored biologically. Shared ASB scaffolds established structural relationships between bioactive and commercial chemical space, and the target profiles of these scaffolds were transferred to commercially available analogs of active compounds. This made it possible to derive target hypotheses for more than 37 000 compounds without biological annotations covering more than 1000 different targets. For many molecules, alternative target assignments were available. Target hypotheses for these compounds should be of interest, for example, for hit expansion, acquisition of compounds to design or further extend focused libraries for drug discovery, or testing of expanded analog series on different targets. They can also be used to search for analogs and complement compound series during target-directed optimization. Therefore, all of the commercial molecules with new target hypotheses as well as key scaffolds identified in our analysis and their target profiles are made freely available.

Drug promiscuity and polypharmacology are much discussed topics in pharmaceutical research. Experimentally, promiscuity can be studied by profiling of compounds on arrays of targets. Computationally, promiscuity rates can be estimated by mining of compound activity data. In this study, we have assessed drug promiscuity over time by systematically collecting activity records for approved drugs. For 518 diverse drugs, promiscuity rates were determined over different time intervals. Significant differences between the number of reported drug targets and the promiscuity rates derived from activity records were frequently observed. On the basis of high-confidence activity data, an increase in average promiscuity rates from 1.5 to 3.2 targets per drug was detected between 2000 and 2014. These promiscuity rates are lower than often assumed. When the stringency of data selection criteria was reduced in subsequent steps, non-realistic increases in promiscuity rates from ~6 targets per drug in 2000 to more than 28 targets were obtained. Hence, estimates of drug promiscuity significantly differ depending on the stringency with which target annotations and activity data are considered.

Providing a large and freely available in silico collection of analog series-based (ASB) scaffolds for computational design and medicinal chemistry applications.

The 'big data' concept plays an increasingly important role in many scientific fields. Big data involves more than unprecedentedly large volumes of data that become available. Different criteria characterizing big data must be carefully considered in computational data mining, as we discuss herein focusing on medicinal chemistry. This is a scientific discipline where big data is beginning to emerge and provide new opportunities. For example, the ability of many drugs to specifically interact with multiple targets, termed promiscuity, forms the molecular basis of polypharmacology, a hot topic in drug discovery. Compound promiscuity analysis is an area that is much influenced by big data phenomena. Different results are obtained depending on chosen data selection and confidence criteria, as we also demonstrate.

Kinase inhibitors are among the most intensely investigated compounds in medicinal chemistry and drug development. Profiling experiments and kinome screens reveal binding characteristics of kinase inhibitors and lead to better understanding of selectivity and promiscuity patterns. However, only limited amounts of profiling data are publicly available. By contrast, a large body of activity data for inhibitors of human kinases has become available from medicinal chemistry. In this study, we have correlated selectivity assessment of clinical kinase inhibitors from the most comprehensive profiling campaign reported to date with systematic mining of activity data from other sources. The results of our comparative analysis reveal consistency of orthogonal approaches in the study of kinase inhibitor selectivity versus promiscuity and stress the importance of taking alternative data confidence criteria into account. Moreover, it is also shown that there are little if any detectable differences in selectivity between type I and II kinase inhibitors and that inhibitors designated as chemical probes have very different target profiles.

Support vector machines are a popular machine learning method for many classification tasks in biology and chemistry. In addition, the support vector regression (SVR) variant is widely used for numerical property predictions. In chemoinformatics and pharmaceutical research, SVR has become the probably most popular approach for modeling of non-linear structure-activity relationships (SARs) and predicting compound potency values. Herein, we have systematically generated and analyzed SVR prediction models for a variety of compound data sets with different SAR characteristics. Although these SVR models were accurate on the basis of global prediction statistics and not prone to overfitting, they were found to consistently mispredict highly potent compounds. Hence, in regions of local SAR discontinuity, SVR prediction models displayed clear limitations. Compared to observed activity landscapes of compound data sets, landscapes generated on the basis of SVR potency predictions were partly flattened and activity cliff information was lost. Taken together, these findings have implications for practical SVR applications. In particular, prospective SVR-based potency predictions should be considered with caution because artificially low predictions are very likely for highly potent candidate compounds, the most important prediction targets.

Combining computational lead optimization diagnostics with analog design and computational approaches for assessing optimization efforts are discussed and the compound optimization monitor is introduced.

Selectivity of kinase inhibitors, or the lack thereof, continues to be an intensely debated topic in drug discovery research. Especially, type I inhibitors, which represent most of the currently available kinase inhibitors, are often thought to lack selectivity because they target the largely conserved adenosine triphosphate-binding site in kinases. Herein, we present a large-scale analysis of potential selectivity among multikinase inhibitors, covering 141 human kinases and more than 10 000 qualifying compounds. By design, the analysis was focused on type I inhibitors and carried out at the level of systematically generated kinase pairs sharing inhibitors. Kinase pair category- and compound-based selectivity profiles identified in part highly selective inhibitors for many kinases. Sets of inhibitors associated with kinase pairs frequently contained nonselective as well as increasingly selective compounds. Selectivity of inhibitors did not result from gatekeeper residues settings or phylogenetic distance of kinases. Rather, it was most likely attributable to subtle differences between binding regions in kinases. Taken together, the results of our study reveal that many multikinase inhibitors are more selective than one might assume.

Ligand docking into homology models of G-protein-coupled receptors (GPCRs) is a widely used approach in computational compound screening. The generation of "double-hypothetical" models of ligand-target complexes has intrinsic accuracy limitations that further complicate compound ranking and selection compared to those of X-ray structures. Given these uncertainties, we have explored "fuzzy 3D similarity" between hypothetical binding modes of known ligands in homology models and docking poses of database compounds as an alternative to conventional scoring schemes. Therefore, GPCR homology models at varying accuracy levels were generated and used for docking. Increases in recall performance were observed for fuzzy 3D similarity ranking using single or multiple ligand poses compared to that of conventional scoring functions and interaction fingerprints. Fuzzy similarity ranking was also successfully applied to docking into an external model of a GPCR for which no experimental structure is currently available. Taken together, our results indicate that the use of putative ligand poses, albeit approximate at best, increases the odds of identifying active compounds in docking screens of GPCR homology models.

Extending the public knowledge base of activity cliffs (ACs) with new categories of ACs having special structural characteristics.

Difficulties in interpreting machine learning (ML) models and their predictions limit the practical applicability of and confidence in ML in pharmaceutical research. There is a need for agnostic approaches aiding in the interpretation of ML models regardless of their complexity that is also applicable to deep neural network (DNN) architectures and model ensembles. To these ends, the SHapley Additive exPlanations (SHAP) methodology has recently been introduced. The SHAP approach enables the identification and prioritization of features that determine compound classification and activity prediction using any ML model. Herein, we further extend the evaluation of the SHAP methodology by investigating a variant for exact calculation of Shapley values for decision tree methods and systematically compare this variant in compound activity and potency value predictions with the model-independent SHAP method. Moreover, new applications of the SHAP analysis approach are presented including interpretation of DNN models for the generation of multi-target activity profiles and ensemble regression models for potency prediction.

Activity cliffs (ACs) consist of structurally similar compounds with a large difference in potency against their target. Accordingly, ACs introduce discontinuity in structure-activity relationships (SARs) and are a prime source of SAR information. In compound data sets, the vast majority of ACs are formed by differently sized groups of structurally similar compounds with large potency variations. As a consequence, many of these compounds participate in multiple ACs. This coordinated formation of ACs increases their SAR information content compared to ACs considered as individual compound pairs, but complicates AC analysis. In network representations, coordinated ACs give rise to clusters of varying size and topology, which can be interactively and computationally analyzed. While AC networks are indispensable tools to study coordinated ACs, they become difficult to navigate and interpret in the presence of clusters of increasing size and complex topologies. Herein, we introduce reduced network representations that transform AC networks into an easily interpretable format from which SAR information in the form of R-group tables can be readily obtained. The simplified network variant greatly improves the interpretability of large and complex AC networks and substantially supports SAR exploration.

A variety of computational screening methods generate similarity-based compound rankings for hit identification. However, these rankings are difficult to interpret. It is essentially impossible to determine where novel active compounds might be found in database rankings. Thus, compound selection largely depends on intuition and guesswork. Herein, we show that molecular networks can substantially aid in the analysis of similarity-based compound rankings. A series of networks generated for rankings provides visual access to search results and adds chemical neighborhood and context information for reference compounds that are not available in rankings. Network structure is shown to serve as a diagnostic criterion for the likelihood to successfully select active compounds from rankings. In addition, comparison of different networks makes it possible to prioritize alternative similarity measures for search calculations and optimize the enrichment of active compounds in rankings.

In pharmaceutical research, compounds with multitarget activity receive increasing attention. Such promiscuous chemical entities are prime candidates for polypharmacology, but also prone to causing undesired side effects. In addition, understanding the molecular basis and magnitude of multitarget activity is a stimulating topic for exploratory research. Computationally, compound promiscuity can be estimated through large-scale analysis of activity data. To these ends, it is critically important to take data confidence criteria and data consistency across different sources into consideration. Especially the consistency aspect has thus far only been little investigated. Therefore, we have systematically determined activity annotations and profiles of known multitarget ligands (MTLs) on the basis of activity data from different sources. All MTLs used were confirmed by X-ray crystallography of complexes with multiple targets. One of the key questions underlying our analysis has been how MTLs act in biological screens. The results of our analysis revealed significant variations of MTL activity profiles originating from different data sources. Such variations must be carefully considered in promiscuity analysis. Our study raises awareness of these issues and provides guidance for large-scale activity data analysis.

Activity cliffs (ACs) are defined as pairs of structurally similar or analogous active compounds with large potency differences [1]. As such, they provide important information for the exploration of structure-activity relationships (SARs) and chemical optimization. We have introduced a new category of ACs capturing minimal (single-atom) chemical modifications and identified more than 1500 of such ACs in compounds with activity against a variety of target proteins [2]. ACs with single-atom modifications (sam_ACs) include "atom-replacement ACs" (ar_ACs) that contain a single-atom replacement (N to C (N-C), O-C, N-O, or S-O) at a given position and "atom-walk ACs" (aw_ACs), in which two analogs are only distinguished by the position of a single heteroatom (non-carbon atom). For a number of sam_ACs, X-ray structures of complexes between AC targets and AC compounds were identified, which made it possible to explore the formation of sam_ACs on the basis of well-defined ligand-target interactions [2]. Our collection of sam_ACs including associated chemical and X-ray structure information, as described herein, is made freely available.

(1) Background: Compounds with multitarget activity are of interest in basic research to explore molecular foundations of promiscuous binding and in drug discovery as agents eliciting polypharmacological effects. Our study has aimed to systematically identify compounds that form complexes with proteins from distinct classes and compare their bioactive conformations and molecular properties. (2) Methods: A large-scale computational investigation was carried out that combined the analysis of complex X-ray structures, ligand binding modes, compound activity data, and various molecular properties. (3) Results: A total of 515 ligands with multitarget activity were identified that included 70 organic compounds binding to proteins from different classes. These multiclass ligands (MCLs) were often flexible and surprisingly hydrophilic. Moreover, they displayed a wide spectrum of binding modes. In different target structure environments, binding shapes of MCLs were often similar, but also distinct. (4) Conclusions: Combined structural and activity data analysis identified compounds with activity against proteins with distinct structures and functions. MCLs were found to have greatly varying shape similarity when binding to different protein classes. Hence, there were no apparent canonical binding shapes indicating multitarget activity. Rather, conformational versatility characterized MCL binding.

The compound optimization monitor (COMO) approach was originally developed as a diagnostic approach to aid in evaluating development stages of analog series and progress made during lead optimization. COMO uses virtual analog populations for the assessment of chemical saturation of analog series and has been further developed to bridge between optimization diagnostics and compound design. Herein, we discuss key methodological features of COMO in its scientific context and present a deep learning extension of COMO for generative molecular design, leading to the introduction of DeepCOMO. Applications on exemplary analog series are reported to illustrate the entire DeepCOMO repertoire, ranging from chemical saturation and structure-activity relationship progression diagnostics to the evaluation of different analog design strategies and prioritization of virtual candidates for optimization efforts, taking into account the development stage of individual analog series.

Currently, there is a high level of interest in deep learning and multitask learning in many scientific fields including the life sciences and chemistry. Herein, we investigate the performance of multitask deep neural networks (MT-DNNs) compared to random forest (RF) classification, a standard method in machine learning, in predicting compound profiling experiments. Predictions were carried out on a large profiling matrix extracted from biological screening data. For model building, submatrices with varying data density of 5-100% were generated to investigate the influence of data sparseness on prediction performance. MT-DNN models were directly compared to RF models, and control calculations were also carried out using single-task DNNs (ST-DNNs). On the basis of compound recall, the performance of ST-DNN was consistently lower than that of the other methods. Compared to RF, MT-DNN models only yielded better prediction performance for individual assays in the profiling matrix when training data were very sparse. However, when the matrix density increased to at least 25-45%, per-assay RF models met or partly exceeded the prediction performance of MT-DNN models. When the average performances of RF and MT-DNN over the grid of all targets were compared, MT-DNN was slightly superior to RF, which was a likely consequence of multitask learning. Overall, there was no consistent advantage of MT-DNN over standard RF classification in predicting the results of compound profiling assays under varying conditions. In the presence of very sparse training data, prediction performance was limited. Under these challenging conditions, MT-DNN was the preferred approach. When more training data became available and prediction performance increased, RF performance was not inferior to MT-DNN.

Machine learning is widely applied in drug discovery research to predict molecular properties and aid in the identification of active compounds. Herein, we introduce a new approach that uses model-internal information from compound activity predictions to uncover relationships between target proteins. On the basis of a large-scale analysis generating and comparing machine learning models for more than 200 proteins, feature importance correlation analysis is shown to detect similar compound binding characteristics. Furthermore, rather unexpectedly, the analysis also reveals functional relationships between proteins that are independent of active compounds and binding characteristics. Feature importance correlation analysis does not depend on specific representations, algorithms, or metrics and is generally applicable as long as predictive models can be derived. Moreover, the approach does not require or involve explainable or interpretable machine learning, but only access to feature weights or importance values. On the basis of our findings, the approach represents a new facet of machine learning in drug discovery with potential for practical applications.